2015 - Referrals Peer Screening and Enforcement in Consumer Credit

American Economic Journal: Microeconomics 2015, 7(3): 174–204
http://dx.doi.org/10.1257/mic.20130234
Referrals: Peer Screening and Enforcement

in a Consumer Credit Field Experiment†
By Gharad Bryan, Dean Karlan, and Jonathan Zinman*
Empirical evidence on peer intermediation lags behind both theory

and practice in which lenders use peers to mitigate adverse selection
and moral hazard. Using a referral incentive under individual lia-
bility, we develop a two-stage field experiment that permits separate
identification of peer screening and enforcement. Our key contribu-
tion is to allow for borrower heterogeneity in both ex ante repayment
type and ex post susceptibility to social pressure. Our method allows
identification of selection on repayment likelihood, selection on sus-
ceptibility to social pressure, and loan enforcement. Implementing
our method in South Africa we find no evidence of screening but
large enforcement effects. (JEL D14, D82, G21, O12, O16)
E conomic theory assigns credit market failure a central role in explaining pov-
erty and underdevelopment. Borrowing constraints reduce efficiency, increase
inequality and can lead to poverty traps (Banerjee and Newman 1993; Galor and
Zeira 1993). Credit rationing also appears to be empirically important. Making use
of experimental or quasi-experimental supply shocks, several recent papers estimate
a large unmet demand for additional credit from consumers, microenterprises, and
small and medium enterprises.1 These studies, coupled with a literature that often
finds high returns to capital (e.g., de Mel, McKenzie, and Woodruff 2008), lend cre-
dence to policy and programmatic efforts to relax borrowing constraints.
But how should one go about relaxing borrowing constraints? Information asym-
metries, including ex ante selection and ex post incentive and enforcement prob-
lems, are often invoked as the root causes of borrowing constraints in theory (Stiglitz
* Bryan: London School of Economics, Houghton Street, London, WC2A 2AE and Innovations for Poverty
Action (e-mail: g.t.bryan@lse.ac.uk); Karlan: Yale University, P.O. Box 208269, New Haven, CT 06520,
Massachusetts Institute of Technology (MIT) Jameel Poverty Action Lab, Innovations for Poverty Action, and
National Bureau of Economic Research (NBER) (e-mail: dean.karlan@yale.edu); Zinman: Department of
Economics, Dartmouth College, 314 Rockefeller Hall, Hanover, NH 03755, MIT Jameel Poverty Action Lab,
Innovations for Poverty Action, and NBER (e-mail: jzinman@dartmouth.edu). The authors would like to thank
Manfred Kuhn and the employees of Opportunity Finance, Luke Crowley, Jon de Quidt, and seminar partici-
pants at Yale University, Northeast Universities Development Consortium Conference (NEUDC), University of
Warwick, Hebrew University, Stanford University Graduate School of Business, Wharton School of the University
of Pennsylvania, The NBER Summer Institute, and The Cambridge conference on consumer credit and bankruptcy.
We would also like to thank The Bill and Melinda Gates Foundation for funding. This paper formed the third chap-
ter of Gharad Bryan’s dissertation—he would like to thank The Kauffman Foundation for financial support. All
errors are, of course, our own.
†
Go to http://dx.doi.org/10.1257/mic.20130234 to visit the article page for additional materials and author
disclosure statement(s) or to comment in the online discussion forum.
1
For consumers see, e.g., Karlan and Zinman (2010), for microenterprises see, e.g., Banerjee (2013) and the
references therein, and for SMEs see, e.g., Banerjee and Duflo (2014).
174
VOL. 7 NO. 3 BRYAN ET AL.: PEER SCREENING AND ENFORCEMENT 175
and Weiss 1981) and practice (Armendáriz and Morduch 2010). If this is indeed the
case, contracts that alleviate asymmetric information problems provide a route to
greater credit market efficiency.
One popular approach to tackling asymmetric information is based on the pre-
sumption that a borrower’s peers can be harnessed to provide information or enforce-
ment that is unavailable to (or more costly for) the lender.2 Peer intermediation has
been fleshed out over several hundred years of lending practice and can be seen in
a range of guises including credit cooperatives, credit unions, rotating savings and
credit associations, and microlenders such as the Grameen Bank. Peer intermediation
has also been analyzed in a large theoretical literature on optimal mechanism design
in the face of different asymmetric information problems (e.g., Varian 1990; Stiglitz
1990; Banerjee, Besley, and Guinnane 1994; Besley and Coate 1995; Ghatak 1999;
Ghatak and Guinnane 1999; Rai and Sjöström 2004; and Bond and Rai 2008).
Empirical work on peer contracting has, however, lagged behind both theory and
practice. Yet we need empirical tests of theories not just to inform and build on our
theoretical models of credit markets, but for practical concerns to inform policy.
Specifically, with more precise information on the absolute and relative importance
and leverage of screening and enforcement, one can design better contracts that
improve credit market efficiency. For example, if peers are able to provide high qual-
ity screening but weak enforcement, a mechanism that uses joint liability to select
clients for a first loan but then moves to individual liability would harness much of
the screening benefit—à la Ghatak (1999)—but minimize the possible negatives of
joint liability lending such as tipping into strategic default (Besley and Coate 1995)
and suboptimal risk-taking (Fischer 2013 and Giné, Goldberg, and Yang 2012).
But if enforcement is largely responsible for the success of peer schemes, then the
appropriateness of peer mechanisms will decrease as ex post enforcement is strength-
ened through debt collection, legal remedies, or better verification of identities.
We formalize and implement a field experiment design that separately identi-
fies peer screening and peer enforcement effects under weaker assumptions than
previous work on asymmetric information. Our key innovations concern the inter-
action between hidden information and hidden action. Our framework allows bor-
rowers to differ both with respect to their ability to repay (their repayment type)
and their susceptibility to social pressure (their malleability type). Heterogeneity
with respect to malleability raises the possibility of selection on malleability—that
peers select based on their ability to enforce repayment—as well as the possibility
that repayment type is correlated with malleability type. We show that these issues
complicate the identification of selection on repayment type—the type of selection
usually associated with peer mechanisms following Ghatak (1999)—but that our
experimental design overcomes these difficulties. The possibility that estimates of
the extent of hidden information (that can be remedied by peer screening) depend
on the extent of hidden action effects (that can be remedied by enforcement) occurs
in many contract design settings. Relative to other papers that have attempted to
separate selection and enforcement in a variety of settings (e.g., Karlan and Zinman
2
Throughout the paper we use the term enforcement to refer to a set of actions that might include monitoring,
peer pressure, mutual insurance, or direct peer assistance.
176 AMERICAN ECONOMIC JOURNAL: MICROECONOMICS AUGUST 2015
2009 and Einav et al. 2013) our experimental design allows us to identify selec-
tion without having to assume a zero correlation between ex ante type and ex post
responsiveness to incentives. Estimates of the extent of selection on malleability are
also important for contract design. Positive selection on malleability implies that the
success of peer enforcement will depend on the amount of selection induced by the
contract. Further, under some reasonable assumptions, selection on repayment type
and selection on malleability will be substitutes, opening up the possibility that the
extent of selection on repayment type will be determined by the lender’s choice of
enforcement technology.
Since theory, practice, and empirical work all suggest that information environ-
ments and the effectiveness of different remedies will vary across settings,3 our
experiment is designed to be portable: simple, and low-cost, to implement in dif-
ferent settings. The hope is that that this will allow the experiment to be used to
tailor contract design to the specifics of particular markets. We demonstrate the
experiment in one setting in conjunction with Opportunity Finance South Africa
(“Opportunity”), a consumer lender located in Kwazulu Natal and a member of the
Opportunity International microfinance network.
Opportunity offered existing clients a 100 rand (US$12) bonus for referring
a new borrower who met particular criteria for Opportunity’s individual liability
loan. Opportunity first randomly divided referrers into one of two ex ante (prior to
referral being made) incentives: referrers in the ex ante approval group were told
that they would receive the bonus if the person they referred was approved for a
loan. Referrers in the ex ante repayment group were told that they would receive
the bonus if the person they referred repaid a loan on time. Referrers in the ex ante
repayment treatment had both an ex ante incentive to refer applicants of good credit
quality (both observable and unobservable to Opportunity), and an ex post incentive
to encourage repayment. Referrers in the ex ante approval group only had an incen-
tive to refer someone they thought would be approved for a loan.
Ex post (i.e., after the referral had been made), Opportunity randomly surprised
some referrers whose referred applications had been approved with an improvement
to their bonus contract.4 Half of the referrers with the ex ante repayment incen-
tive were given their bonuses as soon as the loan was approved, thus removing the
enforcement incentive. Half of referrers given the ex ante approval incentive were
offered an additional bonus if the referred loan was repaid, thus creating an enforce-
ment incentive. Thus, within each of the ex ante groups, half the referrers have an
ex post repayment incentive and half have an ex post approval incentive.5
The design produces four groups of referrers, each with a different combination
of ex ante and ex post incentives, as illustrated in Figure 1. Throughout we denote
the ex ante treatment group (i.e., what the referrer was promised prior to making
the referral) with a lower case letter and the ex post treatment group (i.e., what the
3
For example, Besley and Coate (1995) predicts that the success of peer lending will depend on the strength
of “social collateral” in the community. For empirical evidence compare Giné and Karlan (2014) with Carpena et
al. (2013).
4
Lenders frequently contact borrowers with promotions in this market.
5
The lender also contacted those referrers for whom the contract was not changed and reminded them of the
existence of the contract. This was to dampen any attention or signaling effects.
Ex ante incentive: Bonus promised

Approval Repayment
No enforcement or Screening
Approval screening incentive incentive
Ex post (aA) (rA)
incentive:
Bonus
given Screening and
Enforcement
enforcement
Repayment incentive
incentives
(aR)
(rR)
Figure 1. 2 × 2 Experimental Design—Repayment Incentives
Notes: Bonus given can differ from bonus promised because after offering the initial bonus
ex ante (prior to the referral), the lender later calls each referrer ex post (after the referral is
made) and randomly pleasantly surprises those in the aR and rA cells with a change to their
incentives. Incentives refer to repayment incentives. All referrers have an incentive to refer a
friend that will be approved for a loan.
referrer was given after making the referral) with a capital letter. So someone in
group aR was promised an approval incentive, but also given a repayment bonus.
In Section III we provide a formal model of the referral decision and use it to
show how the design can separate peer enforcement from selection on repayment
type and selection on malleability. Intuitively, enforcement is measured by compar-
ing repayment rates in different rows and selection by comparing repayment rates
across columns. But selection on what? Two additional comparisons permit separate
identification of selection on repayment type and selection on enforcement mallea-
bility. First, comparing repayment rates across columns only in the top row measures
selection in the absence of an enforcement incentive. This comparison reveals the
extent of selection on repayment type even if repayment type and malleability are
correlated. Second, the design produces two estimates of the size of the enforcement
effect, one with a selection incentive (comparing rows conditional on being in the
right column) and one without (comparing rows conditional on being in the left col-
umn). Comparing these two estimates gives us a measure of selection on malleability.
We also show when our experiment can determine whether or not referrers have
information about repayment type. This is a slightly more general question—rather
than asking whether the experiment induced selection, we ask whether the group of
referrers has information. (Referrers might have information but not reveal it if our
mechanism was ineffective.) Guided by an extension to our model (Appendix B),
we show in the empirical section that our repayment incentive induced referrers
to substitute away from referring family toward referring friends. If this substitu-
tion increases the cost of making the referral, and selection on malleability is not
too strong, then we can infer that referrers at least attempted to refer high repay-
ment types. If this attempt was unsuccessful—if we find no evidence of selection
on repayment type—we infer that referrers have a (noisy) belief that they have
information about repayment type, but that in practice this information was already
captured by the lender.
Results from our demonstration show strong enforcement effects. By comparing

repayment rates across ex post incentives holding the ex ante incentive constant,
we find that the small bonus (100 rand is equal to about 2 percent of the average
referrer’s gross monthly income and 3 percent of the average loan size) decreased
default from around 20 percent to 10 percent in most specifications.6 The magnitude
of improvement in repayment performance is far above and beyond what referrers
and borrowers could accomplish with side-contracting, and the improvement (and
savings in collection costs) far exceeded the lender’s outlays for bonuses. Indeed,
the lender continues to use the repayment incentive referral postexperiment.7
We do not find strong evidence of selection effects. Comparing repayment rates
across ex ante incentives, holding the ex post incentive fixed, we find no evidence
that peer selection on repayment type improved repayment, although this is an
imprecise zero. Finally, we are unable to reject that there is no selection on mal-
leability in our setting, although this is again an imprecisely estimated zero with
confidence intervals that contain economically meaningful selection effects.
We make four main contributions relative to the existing literature. First, our
experiment introduces peer influence in a simple individual liability setting where
there is limited potential for complicating strategic interactions.8 This makes identi-
fication possible under plausibly weaker assumptions than required in joint liability
settings.9 Second, we use a two-stage experiment to identify selection and enforce-
ment separately, in a peer referral environment.10 The portability of the experimen-
tal design can help build evidence across settings and contexts. Theory predicts
that peer intermediation will have effects that vary with, e.g., the lenders’ ability
to screen and enforce, the social environment, and the heterogeneity of returns.
Tests using this experimental design across such settings can help build a robust and
empirically-validated theoretical framework. Third, as discussed above, we advance
the study of selection on moral hazard by developing a treatment in which there is
no enforcement incentive.11 Fourth, we provide evidence on the extent of selection
and enforcement effects in a particular setting, and highlight the potential for peer
referral contracts to work in practice.
The paper also provides additional evidence on the presence of asymmetric infor-
mation problems in developing-country credit markets. Our enforcement results
imply the existence of moral hazard, similar to the results in Karlan and Zinman
(2009); Giné, Goldberg, and Yang (2012); de Janvry, McIntosh, and Sadoulet
(2010); and Karlan, Morten, and Zinman (2012). Our lack of a strong screening
6
The emphasis of this paper is on designing contracts that encourage repayment so that the lender can relax bor-
rowing constraints. From this perspective the lower default rate is positive. However, it is plausible that overall wel-
fare is decreased if the social pressure is excessive. So we do not attach a welfare interpretation to any of our results.
7
Besides the enforcement effect, another benefit for the lender is that referrers select on observables: referred
borrowers are much more likely to be approved for a loan (55 percent) than nonreferred borrowers (23 percent).
8
See also Klonner and Rai (2010), which finds in a nonexperimental setting that co-signers improve repayment
performance in “organized” (intermediated) rotating savings and credit associations.
9
For example, Ahlin and Townsend (2007) identifies selection effects only under the assumption that their
model correctly captures the strategic situation.
10
Beaman and Magruder (2012) conduct a peer referral experiment in the labor market and use it to show the
presence of selection effects. They do not consider enforcement, nor the equivalent of selection on malleability.
11
Recent work by Gunnsteinsson (2014) on crop insurance provides an example of how nonenforcement incen-
tives can be applied in other product markets.
result may imply that there is little adverse selection in the market, again similar to
findings in the existing literature. However, we emphasize that there are alternative
interpretations for this null result. Our confidence intervals include economically
meaningful selection effects. It could also be that there is in fact selection on hidden
information, but that peers do not have any better access to this hidden information
than the lender.
The remainder of the paper is structured as follows. Section I introduces
Opportunity and the South African microloan market. Section II provides details of
the experiment. Section III outlines a simple model of the referrer’s decision process,
highlighting the conditions under which our experiment separately identifies enforce-
ment and selection. Section IV provides some summary statistics and discusses the
integrity of the randomization. Section V provides our main results. Section VI dis-
cusses a few alternative explanations of the data and Section VII concludes.
I. Market and Lender Overview
Our cooperating lender is a new entrant to the South African consumer micro-
loan market. Opportunity Finance South Africa (Opportunity) is a for-profit, whol-
ly-owned subsidiary of Opportunity International, which has 1.26 million microloan
customers across 24 different countries. Opportunity operates in the state of
Kwazulu Natal, South Africa, and expanded from one branch in Pietermaritzburg to
five branches across the state during our study period (February 2008 through July
2009). Opportunity offers small, high-interest, uncollateralized debt with a fixed
monthly repayment amount. Loans made during our study period averaged around
3,500R (US$400), with a modal (mean) duration of nine (ten) months, and a modal
(mean) monthly percentage rate of 5 percent (4.1 percent). There is a competitive
market for these loans in Kwazulu Natal (see Karlan and Zinman 2010 for a descrip-
tion of a different lender in this market).
Opportunity underwrites applications using a combination of internal and exter-
nal credit scores (South Africa has well-functioning credit bureaus). A necessary
condition for getting a loan is a documented, steady, salaried job. The loans are
not tied to a specific purpose, but borrowers are asked the purpose of the loan and
most report needing the money for paying school fees for their children, attending/
organizing a funeral, or purchasing a durable good.
II. The Experiment
February 2008 through July 2009, Opportunity offered each individual approved
for a loan the opportunity to participate in its new “Refer-A-Friend” program.
Individuals could participate in the program only once. Referrers received a refer-
ral card, which they could give to a friend (the referred).12 The referred earned
R40 (US$5) if she brought in the card and was approved for a loan. The referrer
12
Note that each referrer received only one card. Had we given each referrer many cards we could have
increased power for the repayment experiment, but the lack of scarcity of cards to give out would have made it
costless to make a referral and thus removed the possibility of studying selection.
If the friend is approved for a loan
If the friend successfully repays a loan
Figure 2. Referral Cards
could earn R100 (US$12)13 for referring someone who was subsequently approved
for and/or repaid a loan, depending on the referrer’s incentive contract.
Opportunity first randomly assigned referrers to one of two ex ante incentive
contracts, corresponding to two different referral cards. Referrers given an ex ante
approval incentive would be paid only if the referred was approved for a loan.14
Referrers given the ex ante repayment incentive would be paid only if the referred
successfully repaid a loan.15 Figure 2 shows examples of the referral cards, the top
card was given to referrers in the ex ante approval group and the bottom card to
those in the ex ante repayment group.
Among the set of referrers whose referred friends were approved for a loan,
Opportunity randomly selected half to be surprised with an ex post incentive change.
Among referrers who had been given the ex ante approval incentive, half were
assigned to receive an additional ex post repayment incentive.16 Opportunity phoned
referrers in the additional-incentive arm and told them that, in addition to the R100
approval bonus, they would receive an additional R100 if the referred repaid the
13
The bonus for the referrer was initially R60 but was changed to R100 in July 2008 at the request of the lender.
The inclusion of this as a control makes no difference in any of our results.
14
Because the bonus was conditional on approval there is the possibility that some referred clients were
approved for a loan and then did not take out a loan. This does not occur in the data.
15
Successful repayment was defined as having no money owing on the date of maturity of the loan.
16
This treatment design generates an additional difference between the treatment group and those receiving
only the ex ante incentive: the referrer receives two payments. Each payment is equivalent to about a day’s wage.
As long as this does not generate an income effect, it should not affect the interpretation of our results.
loan. Opportunity also phoned the other half of referrers in the ex ante approval
incentive group—those who did not get the additional ex post repayment incen-
tive—with a reminder to pick up their R100 bonus. The phone call did not provide
any new information on the incentive contract to this second arm, but we wanted
every referrer to get a call from Opportunity in case the personalized contact from
the lender had some effect. Importantly, while Opportunity was aware of the assign-
ment to treatment, no change was made to the intensity of enforcement in response
to the incentives, and individual staff members responsible for enforcement action
were not directly aware of the treatment assignment.
Among referrers who had been given the ex ante repayment incentive, half
of the referrers were assigned to have the ex post repayment incentive removed.
Opportunity phoned referrers in this arm, told them that they would be paid R100
now instead of conditional on loan repayment, and explained that this was the extent
of the referrer’s bonus eligibility (e.g., that the referrer would not receive an addi-
tional R100 if the loan was repaid). The other half of referrers who had been given
the ex ante repayment incentive were assigned to continue with an ex post repay-
ment incentive. Opportunity phoned these referrers with a reminder that they would
receive a bonus if the loan was repaid.
Figure 1 summarizes the randomization and the incentives that the referrers face.
Intuitively, any effect of peer screening can be identified by comparing the arms
with and without an ex ante repayment incentive, holding constant the ex post incen-
tive. Similarly, any effect of peer enforcement can be identified by comparing the
arms with and without an ex post repayment incentive, holding constant the ex ante
incentive. Two additional comparisons permit separate identification of selection on
repayment type and selection on enforcement malleability. First, comparing repay-
ment rates across columns only in the top row measures selection in the absence of
an enforcement incentive. This comparison reveals the extent of selection on repay-
ment type even if repayment type and malleability are correlated. Second, the design
produces two estimates of the size of the enforcement effect, one with a selection
incentive (comparing rows conditional on being in the right column) and one with-
out (comparing rows conditional on being in the left column). Comparing these two
estimates gives us a measure of selection on malleability.
III. Identification
In this section we provide a stylized model of the referral decision to help clarify
the interpretation of our experiment. We provide the simplest model possible to
discuss the interaction between selection on repayment type, selection on malle-
ability, and enforcement. The purpose is not to provide new theory or derive new
predictions, but rather to formalize what can be identified by the experimental treat-
ments. In particular, we clarify the assumptions needed to interpret the experiment
as separating out selection and enforcement effects. We outline the basic model
in Section IIIA. Section IIIB then relates the model to the data collected in the
experiment. Section IIIC interprets the different repayment rates induced by the
experiment, giving formal definitions of selection on repayment type and selection
on malleability, and then discussing how these are identified by the experiment.
Appendix B analyzes the extent to which the experiment allows us to determine

whether referrers have information about their peers’ repayment types, even if the
lender also has this information.
A. The Referral Decision
A market consists of N potential borrowers (referreds), each of whom is charac-

terized by four parameters:
(i) θ ∈ {θ L, θ H} : a repayment type;
(ii) σ ∈ {σ L, σ H} : a malleability type;
(iii) γ ∈ {γ L, γ H} : the probability that the individual is approved for a loan; and
_
(iv) λ ∈ (0, λ ] : the net cost of recruiting the individual to the referral scheme,
including all social and search costs and/or benefits, but excluding the refer-
ral bonus.
We assume that these characteristics are drawn from a population distribution F , and
allow for characteristics to be correlated. In particular, it may be that repayment type
is correlated with malleability; we discuss below how this would affect identifica-
tion. We concentrate on the case with binary types for ease of exposition.17 Figure 3
provides a simple diagram showing the potential distribution of repayment, mallea-
bility, and approval types. It will be helpful in the discussion below.18
We assume that referrer j knows the cost of referral λij19 and holds beliefs
( )
ˆθij, σˆ ij, γîj about each potential referred i. We further assume that each of these
beliefs is binary. So, for example, θˆ ∈ {θˆ L, θˆ H} with the interpretation that θîj = θˆ H
implies that referrer j believes potential referred i to be a high repayment type.
The timing of the referral decision is as follows:
r Referrer j chooses a referred i and pays a cost λij (or chooses not to make a
referral);
r The referred is approved for a loan with probability γi ;
r The referrer applies social pressure e to encourage repayment and incurs cost
c(e) , where c is a strictly convex cost function;
r The loan is repaid with probability θi + σije.20
17
Allowing for more types does not alter the thrust of the arguments. However, we comment in the text where
allowing for a richer type space would lead to a slightly different interpretation.
18
We leave out correlation with λ as this is assumed to be continuous and it leads to a very complicated diagram.
19
This last assumption seems reasonable given that we assume the cost is paid prior to the referral. Beliefs could
be derived from a more primitive model in which the referrer has a prior belief about types, receives a signal and
uses Bayes’ rule to form the beliefs ( θîj, σˆ ij, γîj).
20
σ is potentially specific to the referrer-referred match, and we allow for that in our notation.
γH θH
X
W
γ L, θ L, σL
σH
Figure 3. Venn Diagram Showing the Distribution of Potential Referreds

by Repayment Type (θ), Malleability Type (σ), and Approval Type (γ)
We assume that {θ H, θ L} and {σ H, σ L} are such that the probability of repay-

ment always lies between zero and one given an optimal effort choice. We could
allow for a more complicated repayment function without altering our discussion.
An advantage of the linear form is that σ encodes all the information needed to
determine the impact of enforcement effort.
In making the referral decision, a referrer takes into account the possibility that she
will exert pressure on the referred to repay the loan. We therefore consider the decision
backward, first analyzing the effort decision and then returning to the referral decision.
Consider first the effort decision. If the referrer is in the ex ante approval treat-
ment she knows she will exert no effort. If the referrer is in the ex ante repayment
treatment she foresees choosing her level of effort to solve
max {BR(θˆ + σˆ e) − c(e)},

e
where BR is the bonus payment in the repayment treatment. We denote the unique
maximizer e ∗(σˆ ), which depends on θˆ only through a correlation between σˆ and θˆ.
For a borrower of type (θ, σ) perceived to be of type (θˆ, σˆ ) the repayment proba-
bility is θ + σe ∗(σˆ ) .
Now consider the referral decision. Given the effort decision discussed above,
referrer j in the ex ante approval group chooses whom to refer by solving
(1) Uj(A) ≡ max {γî BA − λij},

i∈Nj
where BA is the bonus paid in approval treatment. Because the bonus in the ex ante
approval treatment is expected to arrive earlier in time (immediately after approval)
BA ≥ BR. In the ex ante repayment group referrer j solves
Uj(R) ≡ max {γî (BR(θˆ + σˆ e ∗(σˆ )) − c(e ∗(σˆ ))) − λij} .

(2)
i∈Nj
Equation (2) makes clear that the selection decision depends on both the belief
about repayment type and the belief about malleability. Finally, the model is com-
pleted by noting that referrer j makes a referral in the ex ante repayment treatment if
Uj(R) > 0 and in the approval treatment if Uj(A) > 0 .
B. Map from Model to Data
Our analysis focuses on the repayment behavior of individuals referred under dif-
ferent referral incentives. We denote the ex ante (prior to referral being made) treat-
ment assignment with a lower case letter and the ex post (after the referral is made)
assignment with an upper case letter. So, for example, E(λ | aR) is the expectation of
λ conditional on the referrer being in the ex ante approval group and ex post repay-
ment group, while E(λ | a) is the expectation of λ conditional on being in the ex ante
approval group, unconditional on ex post treatment assignment.
We have two potential measures of selection. Define S(A) as the difference
in repayment probabilities across ex ante treatments conditional on being in the
ex post approval treatment (where there is no enforcement incentive). S(A) is given
by the expression
S(A) = E(θ | rA & app) − E(θ | aA & app),
where app implies that the observation is conditional on the referred individual
being approved for a loan, and expectations are taken over the population of refer-
rers. We can also define S(R) as the same comparison conditional on being in the
ex post repayment treatment:
S(R) = E(θ + σe ∗(σˆ ) | rR & app) − E(θ + σe ∗(σˆ ) | aR & app).
We also have two potential measures of enforcement. We define
N(r) = E(θ + σe ∗(σˆ ) | rR & app) − E(θ | rA & app)
as the difference in repayment rates across enforcement treatments conditional on

the ex ante repayment incentive. We can also define the equivalent measure condi-
tional on being in the ex ante approval treatment
N(a) = E(θ + σe ∗(σˆ ) | aR & app) − E(θ | aA & app).
C. Identifying Selection and Enforcement
The identification of enforcement effects is straightforward if the ex post treatment

does not affect the referral or approval decision. This is a reasonable assumption,
and a testable one (Section VB), given that the referral and approval decisions occur
without the referrer or lender staff knowing the ex post treatment assignment. Under
this assumption, repayment type cancels out of the expressions for both N(a) and
N(r) so that N(a) = E(σe ∗(σˆ ) | aR & app) and N(r) = E(σe ∗(σˆ ) | rR & app).
Therefore each of the expressions provides an estimate of the extent to which social
pressure can encourage loan repayment: our experiment generates two estimates of
the peer enforcement effect. These estimates may differ if malleability differs across
the ex ante treatments.
We now turn to identification of selection, focusing first on repayment type and
then on malleability.
DEFINITION 1 [Selection on Repayment Type]: We say that our repayment incen-

tive induced selection on repayment type if
E(θ | r & app) > E(θ | a & app).
That is, on average, the set of clients referred under the ex ante repayment incentive,
and subsequently approved for a loan, are more likely to repay than those referred
under the ex ante approval incentive.
In terms of Figure 3, the probability of repayment conditional on approval is pro-

portional to the number of referreds in areas X and Z (i.e., those that will be approved
and have a high repayment type) relative to area W and Y (i.e., those that will be
approved, but have a low repayment type). We denote this ratio XZ/WY. There is
selection on repayment type if XZ/WY is larger with the ex ante repayment incentive
than with the ex ante approval incentive. We use average repayment rates because they
are relevant to the firm’s profitability and, therefore, its contract design choice.21
We use S(A) to identify selection on repayment type, keeping in mind two issues.
One is whether S(A) identifies repayment type per se (Definition 1), given that refer-
rers in the rA group have incentives to refer borrowers with relatively high malle-
ability and referral cost. Figure 3 illustrates the issue with respect to malleability:
the ex ante repayment incentive may not only change the ratio XZ/WY , but also
the ratios X/Z and W/Y. The key assumption under which S(A) identifies selection
on repayment type is that referrers in treatment group rA do not exert enforcement
effort: given this assumption, the repayment probability is given by θ H for referrers in
areas X and Z , and θ L for those in areas W and Z , and the difference XZ/WZ captures
differences in repayment type.22 Note that S(A) itself does not reveal anything about
other characteristics of referreds besides repayment type. Below we discuss how
we can use S(A) in conjunction with S(R) to identify malleability. The second issue
21
We condition on approval because the approval decision is necessary in credit (and insurance, and labor)
markets. But we need not condition on approval: the extent of selection on repayment type under our definition
will be proportional to the extent of selection on repayment type under the unconditional definition, so long as we
are willing to assume that the probability of approval is strictly increasing in repayment type. This seems like a
reasonable assumption in most settings.
22
Repayment probability is defined as θ + σe as discussed above. As a consequence when e = 0 the repay-
ment probability is θ.
concerns the referrer’s intent. Suppose that referrers make referrals solely on the
basis of malleability type, but malleability type happens to be positively correlated
with repayment type. Definition 1 includes this kind of incidental selection; even
though it is unintentional on the referrer’s part, it is relevant for the lender and the
lender may intentionally design a contract to induce this kind of selection. The ques-
tion of referrer intent does highlight the distinction between identifying selection
and identifying whether referrers have information per se about repayment type. We
pursue this distinction in Appendix B.
Turning to S(R) , we now clarify that this comparison only identifies selection on
repayment type under stronger assumptions. Figure 3 illustrates that WX/YZ may
be smaller in the ex ante repayment group indicating that this group has more high
malleability types. Although this would not affect S(A) because e = 0 , it would
affect S(R) where e ≠ 0 , both in theory and as shown empirically below. If the
effectiveness or extent of social pressure depends on the repayment type (in our
model this would occur through a correlation between repayment type and malle-
ability type) then S(R) could overestimate or underestimate the extent of selection
on repayment type. Moreover, theory does not pin down the sign of the correlation
between repayment type and malleability. One intuition suggests that there is less
scope for social pressure on those who are predisposed to repay: corr(σˆ , θˆ) < 0 and
corr(σ, θˆ) < 0. So in this case high types would be less malleable. Social pressure
might even be counterproductive if high repayment types are intrinsically motivated
and external pressure crowds out intrinsic motivation (e.g., Gneezy and Rustichini
2000; Benabou and Tirole 2003; and Besley and Ghatak 2005). A second intu-
ition, however, suggests that high types will be more malleable; e.g., they already
care the most about repaying and hence will also care most about how they are
viewed by their peers. If this is the case we would expect that corr(σˆ , θˆ) > 0 and
corr(σ, θˆ) > 0.23
We refer to this issue as the comparability problem because S(R) does not pro-
vide an apples-to-apples comparison: the two groups potentially differ in the extent
of social pressure and malleability. The comparability problem occurs in settings
where there is both hidden information, and hidden action that could depend on the
extent of hidden information.
We now consider selection on malleability.
DEFINITION 2 [Selection on Malleability]: We say that our incentive induced pos-

itive (negative) selection on malleability if
E(σe ∗(σˆ ) | rR & app) > (<)E(σe ∗(σˆ ) | aR & app).
That is, on average the set of clients referred under the ex ante repayment incen-
tive, and subsequently approved for a loan, are more (less) susceptible to the
23
In the more familiar setting of moral hazard and adverse selection in the credit market, Karlan and Zinman
(2009) provide a simple model that implies high types put more effort into ensuring project success. If that model
is true, correlation between hidden information and the extent of hidden action would tend to lead to an underesti-
mate of the extent of selection on risk type. Even in that case, however, it is possible to tweak the simple model and
reverse the comparative static. The extent and direction of these interaction effects is, therefore, an empirical matter.
chosen social pressure of their referrer than those referred under the ex ante
approval incentive.
In terms of Figure 3 we say that there is selection on malleability if YZ/WX

depends on the treatment assignment.24 We condition on the optimal choice of
social pressure in the ex post repayment treatment. An alternative would be to define
positive (negative) selection on malleability to be
E(σ | r & app) > (<)E(σ | a & app).
This may seem a more natural definition, but a more malleable set of referrers is not
helpful to the lender if the referrers do not in fact put the appropriate pressure on
these referrers.
Again assuming that the ex post treatment does not affect the referral or approval
decision we have
S(R) − S(A) = E(σe ∗(σˆ ) | rR & app) − E(σe ∗(σˆ ) | aR & app).
Therefore, S(R) − S(A) > (<)0 implies positive (negative) selection on

malleability.25
In Appendix B we use the model to show that if E(λ | r) > E(λ | a) , so that the
cost of referrals in the ex ante repayment treatment is higher than in the ex ante
approval treatment, and S(A) = 0 then we can infer that referrers do not have infor-
mation about repayment type, conditional on the information used by the lender for
approval. The intuition is simple: if the referral is more costly in the ex ante repay-
ment group, then we can infer that referrers tried to choose someone who is likely
to repay. But if S(A) = 0 , there is no effect of the referrers’ selection, and we can
infer that referrers failed in their attempt to screen on repayment type. Appendix B
clarifies some additional assumptions required to complete the argument.
IV. Data
A. Summary Statistics
Table 1 provides a summary of the characteristics of potential referrers over the

period in which the experiment was run.
24
As above, we condition on approval and look at average repayment rates as we believe that this is the most
appropriate measure when considering how to design contracts.
25
There is a meaningful distinction between intentional and incidental selection on malleability if referrers are
not forward-looking or do not actually have a good signal of malleability type. Although these contingencies seem
unlikely to arise, we note two points. First, it is the possibility of incidental selection on malleability—in particular
of high repayment types being less malleable—that leads to the possibility of negative selection on malleability.
Second, observing S(A) < 0 and S(R) > 0 is only possible if there is positive intentional selection on malleability,
and if selection on repayment type and selection on malleability are substitutes.
Table 1—Demographic Variables of All Borrowers during Experiment
Mean Median SD
Female 0.421 — 0.494
Age 37.786 36.000 10.797
High school education 0.633 — 0.482
Disposable income 1,746 1,254 1,705
Requested amount 5,023 3,000 6,592
Requested term (months) 10.735 9 7.257
Observations 4,408
Notes: Disposable income is monthly income remaining after rent, debt repayments, and recur-
ring obligations. It is measured in rand. An individual has a high school education if they have
matriculated or gone on to tertiary education.
B. Integrity of the Randomization
Opportunity handed out 4,408 referral cards to potential referrers—borrowers

approved for new loans—during the study period.26 Table 2 presents regressions of
treatment assignment on referrer characteristics at baseline. If the randomization is
valid, we would expect these characteristics to be uncorrelated with treatment. In all
cases an F-test of the restriction that the coefficients are jointly zero fails to reject
at the usual significance levels. Further, most individual coefficients are not statis-
tically different from zero and the total number of significant coefficients is in line
with what we would expect to see by chance. Below we also show that our treatment
effect estimates are robust to including controls for referrer baseline characteristics.
Of the 4,408 cards that were handed out, 430 were returned and 245 of these
referred clients were approved for a loan. The surprise nature of the second ran-
domization (i.e., the change in ex post incentives) provides another opportunity to
check the integrity of the experimental implementation. Because the second-stage
assignments were not known to potential referrers ex ante (nor to Opportunity
staff members delivering referral cards), baseline characteristics of those referred
and approved for a loan27 should not differ within the ex post treatment groups.28
Table 3 presents tests of this hypothesis from regressions where the outcome vari-
able is being assigned to the ex post repayment incentive.
Within the group given the ex ante approval incentive (Table 3, column 1), the
F-test shows that the baseline coefficients do not significantly predict assignment to
treatment in the joint test. Among the individual tests, only one of the 16 variables is
significant, which is about what one would expect to happen by chance.
Within the group given the ex ante repayment incentive (column 2), a higher
application score (i.e., internal credit score) significantly predicts assignment to the
ex post approval group. Given that application score is a key measure of the observed
credit quality of the applicant, this is troubling. It turns out that Opportunity changed
26
Referred clients were not eligible for a card. This restriction was aimed to reduce learning about the treat-
ments among potential referreds.
27
Results are similar if we do not condition on approval, and instead consider the full sample of 430 referreds.
28
Comparison across the ex ante incentive groups are, however, endogenous. That is, we cannot compare char-
acteristics of those in the ex ante approval groups to those in the ex ante repayment groups because the experiment
aims to generate difference in these characteristics.
Table 2—Testing the Balance of Referrer Characteristics

across Treatments: OLS
Ex ante incentive Approval Repayment

Ex post incentive Approval Repayment Approval Repayment
Female −0.006 0.008 0.005 −0.008
(0.014) (0.014) (0.014) (0.014)
Age 0.000 −0.001 0.000 0.000
(0.001) (0.001) (0.001) (0.001)
High school education −0.027 0.027 −0.007 0.008
(0.023) (0.023) (0.023) (0.023)
Salary earner −0.004 0.022 −0.016 −0.003
(0.016) (0.016) (0.016) (0.016)
Disposable income −0.003 −0.006 0.010* −0.001
(thousands of rand) (0.004) (0.004) (0.004) (0.004)
Application score 0.000 0.000 0.000 0.000
(0.000) (0.000) (0.000) (0.000)
ITC score 0.000 0.000 0.000 0.000
(0.000) (0.000) (0.000) (0.000)
ITC score missing 0.072 0.039 −0.058 −0.053
(0.109) (0.109) (0.109) (0.110)
Requested amount −0.001 0.004* −0.004* 0.000
Requested term 0.002 −0.004* 0.002 0.000
(months) (0.002) (0.002) (0.002) (0.002)
Government worker 0.005 −0.002 0.022 −0.025
(0.031) (0.031) (0.032) (0.032)
Cleaner/builder/miner 0.010 −0.006 0.007 −0.011
(0.031) (0.031) (0.031) (0.031)
Security/mining/transport 0.021 −0.004 0.020 −0.037
(0.033) (0.033) (0.033) (0.033)
Retail worker −0.002 0.008 0.003 −0.009
(0.031) (0.031) (0.032) (0.032)
IT/financial worker 0.010 −0.025 0.014 0.001
(0.035) (0.035) (0.035) (0.035)
Agriculture/manufacturing 0.008 −0.005 0.027 −0.030
−0.029 −0.029 −0.029 −0.029
Constant 0.189 0.224 0.273** 0.315*
(0.118) (0.118) (0.119) (0.120)
F-test of joint significance 0.560 0.930 0.810 0.440

p-value of F-test 0.916 0.533 0.679 0.971
Observations 4,408 4,408 4,408 4,408
Notes: Each column represents a separate OLS regression where the LHS variable is a dummy
taking the value of 1 if the individual is assigned to the particular treatment. Education is a
dummy variable taking on value 1 if the referrer has matriculated. Application score is an inter-
nal credit score. ITC score is an external credit score. Salary monthly is a dummy variable tak-
ing value 1 if the client receives his or her salary monthly.
*** Significant at the 1 percent level.
** Significant at the 5 percent level.
* Significant at the 10 percent level.
its application score in May 2009 (before this time, scores are out of 200, while after
they are out of 800). Only 12 referred clients from the ex ante repayment group were
approved for loans after the change, with 9 coming from the ex post approval group.
This is not out of line with what we would expect from random arrival times, but it
does create a problem in testing orthogonality. Columns 3 and 4 of Table 3 take two
approaches. Column 3 omits the application score, and the p-value for the F-test
of joint significance rises to 0.326 (from 0.077 in column 2). Column 4 drops the
12 post-change observations (so the sample size falls from 120 to 108), and the
p-value is 0.478. We also note that the ITC (external) credit score—which is pro-
vided by a credit bureau and is another key predictor of credit worthiness—is never
predictive of treatment status.
Overall it seems that the randomization was successful. We also show below that
our results are not sensitive to including these baseline characteristics as controls.29
V. Results
A. Who Refers and Who Gets Referred?
Among eligible clients, 430 of 4,408, or 10 percent, made a referral. Making a

referral is not correlated with the ex ante incentive: a regression of the referral rate
on a dummy for ex ante treatment gives a coefficient of −0.002 with a standard
error of 0.040. This suggests that the higher net present value of the bonus in the
ex ante approval group was not an important consideration in the referrers’ calcu-
lations and did not induce greater search. And it bodes well for using the assump-
tion that E(λ | a) < E(λ | r) to help identify whether referrers have information (see
Appendix B).
We also collected data on the relationship between referrers and referreds.
Specifically, the lender asked the referred “How do you know the person that gave
you the voucher?” 426 of our referreds answered this question, and most stated that
the referrer was a relative or work colleague. A small number answered church.
Table 4 shows the results from a multinomial logit of the impact of being in the
ex ante repayment group on the likelihood of referring a coworker or someone from
the same church, relative to referring a family member. The table shows that the
selection incentive leads referrers to substitute away from relatives toward those
they know through church and their work colleagues.
We argue that this finding implies that E(λ | r) > E(λ | a). First, we believe that
it is intuitive that colleagues and co-religionists are more costly to refer than family
members. This is both because colleagues may be harder to approach and because
there may be a positive obligation to refer family members given the R40 bonus
given to the referred. Second, Beaman and Magruder (2012) find similar substitution
toward referring colleagues in a labor market experiment in India. Their experiment
also documents that referred colleagues are more productive workers than referred
family members, which would only occur if colleagues are indeed more costly to
29
We can only control for these differences when studying the enforcement question; when we consider selec-
tion, referred characteristics are endogenous.
Table 3—Testing the Balance of Referred Characteristics

across Ex Post Treatments
Whole App. score Before May

sample excluded 2009
Ex ante incentive Approval Repayment Repayment Repayment
Female 0.041 0.098 0.088 0.147
(0.113) (0.104) (0.107) (0.113)
Age 0.001 0.003 0.004 0.002
(0.005) (0.005) (0.005) (0.006)
High school education 0.131 0.022 −0.005 −0.008
(0.148) (0.164) (0.169) (0.173)
Salary earner 0.029 −0.117 −0.086 −0.106
(0.110) (0.113) (0.116) (0.133)
Disposable income 0.034 0.028 0.037 0.023
Application score/100 0.016 −0.071*** — 0.020
(0.000) (0.000) — (0.004)
ITC score 0.002 0.000 0.000 0.000
(0.001) (0.001) (0.001) (0.001)
ITC score missing 1.197 −0.116 −0.276 −0.077
(0.893) (0.895) (0.922) (0.935)
Requested amount 0.010 −0.027* −0.026 −0.020
Requested term −0.013 0.014 0.010 0.012
(months) (0.015) (0.017) (0.018) (0.019)
Government worker −0.389 0.103 0.023 0.115
(0.266) (0.257) (0.264) (0.273)
Cleaner/builder/miner −0.094 0.098 0.028 0.027
(0.207) (0.211) (0.217) (0.225)
Security/mining/transport −0.330 −0.321 −0.450* −0.355
(0.226) (0.251) (0.255) (0.275)
Retail worker −0.212 −0.105 −0.166 −0.129
(0.203) (0.220) (0.226) (0.231)
IT/financial worker −0.570** 0.495 0.445 0.427
(0.279) (0.533) (0.550) (0.562)
Agriculture/manufacturing −0.222 −0.168 −0.214 −0.199
−0.187 −0.222 −0.229 −0.234
Constant −0.547 0.642 0.692 0.348
(0.923) (0.932) (0.962) (1.059)
F-test of joint significance 0.810 1.640 1.150 0.990

p-value of F-test 0.669 0.077* 0.326 0.478
Observations 123 120 120 108
Notes: Dependent variable is assignment to ex post repayment incentive. Each column rep-
resents a separate OLS regression where the LHS variable is assigned to the particular treat-
ment. Education is a dummy variable taking on value 1 if the referrer has matriculated.
Application score is an internal credit score. ITC score is an external credit score. Salary
monthly is a dummy variable taking value 1 if the client receives his or her salary monthly.
Table 4—The Impact of Ex Ante Treatment Group on Referred Characteristics:

Multinomial Logit, Compared to Referring a Relative
Same church Coworker

Ex ante 1.026** 0.434*
Repayment incentive (0.434) (0.224)
Constant −1.932*** 0.686***

(0.338) (0.148)
Observations 426
Notes: Results from a multinomial logit. Omitted category is referral of a relative. Standard
errors are in parentheses.
refer. Consequently, following our analysis in Appendix B, if we can now show that
S(R) = S(A) and S(A) = 0, we will infer that referrers have no information.
We also find evidence that referred clients are more likely to be approved for a loan.
The approval rate for clients off-the-street is around 23 percent, but for clients referred
through the Refer-A-Friend program the approval rate is around 55 percent. We will
argue below that referrers were not successful in selecting on repayment type. The
observation that referred clients are more likely to be approved is, therefore, consistent
with two interpretations: (i) peers know which of their friends are creditworthy, but
this information duplicates information already held by the lender; and (ii) peers have
correlated credit scores and, because the referrers were all approved borrowers, their
peers are more likely to be approved than an average client.30 These two possibilities
make it hard to give a causal interpretation to the correlation.
B. Measuring Repayment Performance
We identify screening and enforcement effects by comparing the repayment

performance of loans referred by referrers facing different incentives. We have
four different and complementary measures of repayment performance. Each
proxies for the costs a lender bears when borrowers don’t repay (on time), with-
out needing to impose additional assumptions on what the lender’s cost structure
actually is (since in our experience many lenders lack precise data on marginal
costs of collections). First, we have an indicator variable, for all 245 referred
clients, of whether or not the borrower was charged penalty interest for paying
late at any time during the course of the loan. Second, we measure whether the
loan was fully repaid on the date of maturity for the 240 loans that have reached
maturity. Third, for those 240 loans we also calculate the proportion of principal
still owed at maturity date (this value is zero for loans repaid on time, and positive
30
It is also possible that the lender took into account whether the borrower was referred in making its approval
decision. We find this unlikely for two reasons. First, the approval process is quite mechanical. Second, the lender
was genuinely agnostic as to whether the referral system would work. This was one of the main reasons for imple-
menting the experiment.
for loans in arrears). Fourth, Opportunity charges off loans deemed unrecoverable
and has made a chargeoff decision (yes or no) on all but 1 of the 240 loans that
have reached maturity as of this writing.31
Each panel in Table 5 shows the mean of these four loan performance measures,
organized by treatment groups. It also shows the difference in means holding either
the ex ante or ex post incentive fixed. These differences are our key results. Relating
these back to Section IIIB the rightmost cell in the top row of each panel gives S(A) ,
the rightmost cell in the second row gives S(R), the bottom cell in the first column
gives N(a), and the second column gives N(r). Finally, the bottom right cell in each
table gives | S(R) − S(A) |.
C. Enforcement Effects
Table 5 provides eight estimates of the peer enforcement effect, two for each
measure of default. The bottom cell in the first column of each panel shows an esti-
mate of N(a) and the bottom cell in the second column shows an estimate of N(r).
The point estimate for each of the eight differences is negative, suggesting that add-
ing the ex post repayment incentive decreases the incidence of default. Five of the
eight estimates are statistically significant from zero, despite our small sample.
In each case the implied magnitude of the enforcement effect is large; e.g., N(a)
in panel D implies an 11 percentage point reduction in chargeoff likelihood, on a
base of 16 percent. The size of the impacts is also large given the size of the bonus
(R100) and the average loan size. It seems unlikely that these results could be driven
by referrers transferring R100 to the referred to encourage repayment. In all, the
results suggest that the small referral incentive created social pressure that led to
large reductions in default.
It is interesting to ask how the size of the effect compares to the impact of an
incentive given directly to the borrower—rather than to a peer. Although we do not
have the ideal data, we do have one bit of evidence from a similar context with which
to do a back of the envelope calculation. Karlan and Zinman (2009) conducted a
dynamic incentive experiment with a similar, although much larger, South African
lender in 2004. That intervention is somewhat different in that the dynamic incentive
did not come in the form of a cash bonus, but rather in the form of a reduced rate on
a future loan. On average, the dynamic incentive reduced the interest rate on a future
loan by 3.85 percentage points and led to a roughly 2.5 percentage point increase in
likelihood that the current loan was paid on time. This result suggests that to have
a similar impact as our study, a direct incentive would need to be very large—in
the order of a 12 percentage point reduction in the interest rate (effectively making
the interest rate on the next loan zero). This again suggests that at least part of the
enforcement effect in our experiment reflects social pressure, rather than simply the
transfer of cash from the referrer to the borrower.
31
The results do not change qualitatively if we arbitrarily assign this loan as being charged off or not.
Table 5—Key Outcome Variables: Mean Differences across Treatment Groups
Ex ante incentive Ex ante incentive

Approval Repayment Diff. Approval Repayment Diff.
Panel A. Penalty interest charged by lender Panel B. Positive balance owing at maturity
(observations = 245) (observations = 240)
Ex post incentive Ex post incentive
Approval 0.389 0.518 0.129 Approval 0.206 0.226 0.019
(0.064) (0.069) (0.093) (0.054) (0.058) (0.079)
Repayment 0.258 0.272 0.015 Repayment 0.095 0.152 0.056
(0.054) (0.055) (0.077) (0.037) (0.044) (0.058)
Difference −0.132 −0.246*** 0.114 Difference −0.111* −0.075 0.036
(0.083) (0.087) (0.122) (0.064) (0.072) (0.098)
Panel C. Portion of value owing at maturity Panel D. Loan charged off by lender
(observations = 240) (observations = 239)
Ex post incentive Ex post incentive
Approval 0.187 0.257 0.070 Approval 0.155 0.188 0.034
(0.054) (0.076) (0.091) (0.048) (0.054) (0.072)
Repayment 0.076 0.109 0.033 Repayment 0.047 0.092 0.045
(0.039) (0.039) (0.055) (0.027) (0.036) (0.045)
Difference −0.110* −0.147* 0.037 Difference −0.108** −0.096 0.011
(0.066) (0.081) (0.108) (0.054) (0.063) (0.085)
Notes: Penalty interest is charged by the lender if a borrower is late in making an expected payment. A loan is
charged off if the lender deems that there is no probability that it will be repaid. Standard errors are in parenthe-
ses and p-values in square brackets. p-values are for a χ2-test of the hypothesis that the difference-in-differences
is equal to zero. Ex ante incentive is the incentive that the referrer faced when choosing a friend to refer. Ex post
incentive is the incentive that the referrer faced after the loan had been approved. Approval implies the loan had to
be approved in order to earn the bonus, and repayment implies the loan had to be repaid in order to earn the bonus.
D. Selection Effects
Table 5 also provides eight estimates of the peer selection effect, two for each
measure of default. Relating these back to Section IIIB, the rightmost cell in the
top row of each table gives S(A) , and the rightmost cell in the second row gives
S(R). Each of the eight point estimates is statistically insignificant and positive, so
there is no evidence that a small referral incentive induces screening on repayment
type that reduces default. In fact, if we take the point estimates seriously they
imply negative selection on repayment type. This outcome is consistent with our
model if there is strong intentional positive selection on malleability. As discussed
in Section IVC, this could lead to the crowding out of the incentive to select on
repayment type.
The bottom-right cell in each panel of Table 5 estimates the difference-in-
differences (DD) across the two different estimates of the referral incentive effects on
default rates. Recall from Section III that if S(A) = S(R) we can rule out selection
on malleability. A zero estimate of the DD indicates that malleability is uncorrelated
with ex ante repayment type. And indeed none of the four estimates is significantly
different than zero. This suggests that, in our setting, selection on malleability and
the comparability problem are not that important. It bears emphasizing, however,
that these are very imprecisely estimated zeros: each of the four confidence intervals
includes economically large selection on malleability.
The DD estimate and the estimates of the selection effect are also relevant for
determining whether referrers have information. In Section VA we argued that
E(λ | r) > E(λ | a). This assumption, following the analysis of Appendix B, in com-
bination with the findings above that S(A) = 0 and S(R) − S(A) = 0 , suggest
that referrers have no useful information for the lender (again subject to the above
mentioned caveats on the distribution of repayment types and our wide confidence
intervals). While this result should not be over interpreted (it requires quite strong
assumptions) it is worth noting that the alternative approach to measuring whether
people have information—getting people to bet on the loan performance of their
friends—will not allow for the removal of enforcement effects.
E. Improving Precision and Checking Robustness
Under the assumption that malleability is uncorrelated with repayment type,

which is born out by the above DD estimates, our model implies that we can esti-
mate the extent of peer enforcement and selection with greater precision, using
regressions that pool across all four treatment arms:
yi = α + β 1 enforcei + β 2 selecti + ϵi ,
where yi is one of the four measures of default, enforcei is an indicator with value 1
if client i was referred by someone with the ex post repayment incentive, and
selecti is an indicator with value 1 if the client was referred by someone with the
ex ante repayment incentive. Results from this regression (without controls) are
presented in Table 6. For each of the four outcome measures we see a large and
statistically significant reduction in default associated with the enforcement incen-
tive, and a smaller and statistically insignificant increase in default coming from
the selection incentive. These results sharpen the key inferences from the means
comparisons in Table 5: there is a large enforcement effect, and no (or a perverse)
selection effect.
Appendix A shows that these results are robust to various specifications that con-
trol for the baseline characteristics of borrowers or referrers.
VI. Alternative Explanations and Other Considerations
In this section we discuss alternative interpretations of the results.
A. Income Effects
In theory, the enforcement effect could be driven by side-payments from the

referrer to referred that produce an income effect on loan repayment. In practice
this channel seems implausible, for several reasons. First, the bonus was not paid
out until after the loan was repaid, and the borrowers in our sample are liquidity
constrained (as evidenced by the fact that they are borrowing at high rates). Second,
Table 6—Pooled Impact of Selection and Enforcement Treatments on Key Outcome Variables:
OLS without Controls
Outcome measure Penalty interest Not paid on time Portion owing Loan charged off
Enforcement −0.188*** −0.094* −0.129** −0.102**
(0.061) (0.049) (0.054) (0.042)
Selection 0.067 0.039 0.050 0.040
(0.060) (0.047) (0.052) (0.041)
Constant 0.419*** 0.197*** 0.196*** 0.152***
(0.054) (0.045) (0.046) (0.040)
Notes: Penalty interest is charged by the lender if a borrower is late in making an expected payment. A loan is
charged off if the lender deems that there is no probability that it will be repaid. Standard errors are in parentheses.
Ex ante incentive is the incentive that the referrer faced when choosing a friend to refer. Ex post incentive is the
incentive that the referrer faced after the loan had been approved. Approval implies the loan had to be approved in
order to earn the bonus, and repayment implies the loan had to be repaid in order to earn the bonus.
even our smaller point estimates imply default reductions that seem too large (about
R500 on the average loan) to be explained by a small increase in income (maximum
R100). Third, as discussed in Section VC, the enforcement effects here are large in
comparison to effects from bonuses paid directly to the borrower.
B. Signaling
The repayment rates in Table 5 consistently show that the highest default rates
occur for those clients who were in the ex ante repayment group and moved to the
ex post approval group. In this treatment group, Opportunity phoned the referrer
and told her that the bonus would no longer be paid upon repayment. It is possi-
ble that this signaled that the lender was not really interested in repayment. If this
explanation is correct, then our estimate of selection conditional on being in the
ex post approval group would be biased in favor of showing no screening, while our
estimate of enforcement conditional on the ex ante repayment incentive would be
biased in favor of finding an enforcement effect.
Although we cannot completely rule out the lender perversely signalling refer-
rers in the rA arm, three factors suggest that it is not a first-order concern. First,
recall that we can estimate the enforcement effect and selection effect without
relying on the rA arm. Second, recall that our estimates of selection and enforce-
ment effects do not actually differ whether we use the rA arm or not: we do not
find any evidence, statistically speaking, that estimates using the rA arm are
any different. (This inference is of course subject to the caveat that our confi-
dence intervals are wide and hence any null result in this paper does not rule out
large differences). Third, if the signaling story were correct the repayment rates
for the referrer would also be affected, since the referred would only receive the
perverse signal through the referrer (the lender does not directly provide the bor-
rower with any information about referral incentives). But the difference between
the default rates of the “signaled” and the “un-signaled” is not significantly dif-
ferent from zero for any of the four default measures (−0.060 ( p = 0.414),
−0.030 ( p = 0.473), −0.019 ( p = 0.664), and −0.006 ( p = 0.895)): the data
does not support the signaling story. If anything the point estimates suggest that the
“signaled” were better repayers.
C. Impatience
Referrers who were assigned to the ex ante repayment incentive were promised a
bonus that would not be paid until the referrer repaid their loans. One might therefore
expect fewer referrers to make a referral in this treatment group, and/or that those
making referrals would be more patient (and hence be more willing to and effec-
tive at enforcing loans). Either difference could, in principle, create issues for the
identification of screening effects. In practice, such issues do not matter much. First,
the number of referred clients does not differ across the ex ante treatment groups
(99 in the ex ante approval group versus 94 in the ex ante repayment group,
p = 0.516). Second, if referrers in the ex ante repayment group were more patient
and this impacted how much social pressure they placed on their referreds, then we
would expect to see evidence for this in the size of the enforcement effect; i.e., esti-
mates using the ex ante repayment group should show larger effects. As discussed
above, there is no evidence for this, albeit with large confidence intervals.
D. Reputation and Other Considerations
Most borrowers probably see themselves in an ongoing relationship with the

lender. They may, therefore, value their reputation with the lender and this may give
an incentive to enforce the loan even if they did not receive an ex post repayment
incentive. This could have two effects. First, it may weaken our ability to identify
enforcement effects if the proportional effect of reputation on enforcement is larger
in the ex post approval than the ex post repayment arms. This is not a major con-
cern in the current implementation, given that we find large enforcement effects in
spite of any downward bias. Second, it could reduce our ability to identify selection
effects if the effect of reputation on screening is proportionally larger in the ex ante
approval than the ex ante repayment arms. This would occur, for example, if reputa-
tion effects imply that all referrers engage in the maximal amount of selection. But,
the fact that we see referrers move toward referring workmates in response to the
selection incentive suggests that our incentive was effective in inducing an attempt
at screening, so we believe that the reputation effect cannot have been too strong.
Another issue arises when comparing across ex ante incentive holding fixed the
ex post repayment incentive—i.e., comparison S(R). In this case, those in the aR
group have already received a payment from the lender when they consider enforce-
ment, while those that are in the rR group have not. This may mean that those in the
aR group believe more strongly in the lender’s willingness to pay, and hence have
a larger incentive to enforce. This effect will tend to lead us to underestimate any
selection effect when using the comparison S(R). This logic will not, however, affect
the validity of the S(A) comparison as a measure of selection on repayment type, but
will weaken our ability to use the difference S(R) − S(A) as a measure of selection
on malleability.
A related issue arises from the timing difference in the offer of ex ante and ex post
repayment incentives. Referrers in the rR group are notified of their incentive earlier
than those in the aR group and may thus engage in different kinds of enforcement
(e.g., influencing their referred’s project choice). If this makes enforcement more
effective in rR than aR , then S(R) will overestimate selection effects. But the impact
of this potential bias is mitigated by the fact that we can and do use S(A) instead of
S(R) to identify selection on repayment type. Note also that in practice we do not
find statistically significant selection effects, meaning that any overestimation using
S(R) is not producing a false positive, at least in our setting.
VII. Conclusions
We formalize and implement a field experiment design and show it is possible

to identify peer selection and peer enforcement effects separately, even when peers
select on malleability to enforcement pressure. Novel features of the design include
experimental arms without enforcement incentives (see also Gunnsteinsson 2012),
combinations of experimental arms that produce different estimates of selection and
enforcement effects, and a combination that estimates the extent of selection on hid-
den action (malleability). We hope that our analysis of identification will pave the
way for further progress in identifying interactions between hidden information and
hidden action, using multistage experiments as well as other methods.
Our empirical results imply large peer enforcement effects. They also imply peer
selection that is only partly effective: peers refer applicants who are likely to be
approved, but do not end up producing any ex ante information on repayment type
that is useful to the lender on the margin (despite evidence suggesting that they exert
at least some effort to select better repayment types).
Both theory and practice suggest that results may be different elsewhere—e.g.,
in settings that lack a well-functioning credit bureau, or with contracts that use
larger referral bonuses—and hence our experiment is designed to be portable. Our
experiment can be layered on top of any individual liability lending operation, and
could be modified to work with joint liability lending. The fact that referral incen-
tives were quite profitable for the lender in our demonstration could help convince
lenders (and firms in other product markets) that it is worth experimenting with
referral mechanisms.
Appendix A. Robustness to Controls
We now check whether the results are robust to adding controls. We start by esti-
mating the enforcement or screening effect separately using equations of the form:
(A1) yi = αi + β Ti + γ Xi + ϵi ,
where yi is again a measure of default, Ti is a dummy variable that takes on value 1

if i is “treated,” and Xi is a set of controls for either referrer or borrower baseline
Table A1—Enforcement Effects.

The Impact of Ex Post Repayment Incentive within Ex Ante Treatment Group:
OLS with Controls for Referrer Characteristics
Penalty interest Not paid on time Portion owing Charged off

Panel A. Ex ante approval incentive
Ex post approval Left out Left out Left out Left out
Ex post repayment −0.155 −0.193** −0.189** −0.151**
(0.097) (0.084) (0.086) (0.071)
Mean in ex post 0.389 0.206 0.186 0.155
approval (0.064) (0.054) (0.054) (0.047)
Controls All All All All
Panel B. Ex ante repayment incentive

Ex post repayment −0.206* −0.089 −0.129 −0.129
(0.112) (0.087) (0.091) (0.078)
Mean in ex post 0.519 0.226 0.256 0.189
approval (0.069) (0.058) (0.076) (0.054)
Notes: Penalty interest is charged by the lender if a borrower is late in making an expected
payment. A loan is charged off if the lender deems that there is no probability that it will be
repaid. Standard errors are in parentheses. Ex ante incentive is the incentive that the referrer
faced when choosing a friend to refer. Ex post incentive is the incentive that the referrer faced
after the loan had been approved. Approval implies the loan had to be approved in order to
earn the bonus, and repayment implies the loan had to be repaid in order to earn the bonus.
Controls: Female, Age, Disposable Income, Salary Occurrence, Education, Application Score,
ITC Score, Job Type, Requested Loan Amount, Requested Term, Branch, Application Month,
Application Year. All controls are for referrer characteristics. Categorical variables are entered
as fixed effects.
characteristics (these sets of characteristics are highly collinear). When estimating

the enforcement effect here, Ti = 1 if the referrer was given the ex post repayment
incentive. We condition on the ex ante incentive by running regressions separately
for the samples that received the ex ante approval incentive (Tables A1 and A2,
panel A) or the ex ante repayment incentive (panel B). When controlling for the
referred’s application score, we include a dummy variable for whether the client
came in after the change in application score procedure and also interact that term
with the application score. Tables A1 and A2 show that adding controls does not
alter the coefficients appreciably.
To test for selection effects we repeat the above exercise with Ti being an indica-
tor for whether the referrer was given an ex ante repayment incentive. The results
are reported in Table A3. In panel A we restrict the sample to those given the ex post
approval incentive and in panel B we restrict the sample to those given the ex post
repayment incentive. For these regressions we control for referrer characteristics as
the referred characteristics are endogenous. Again, the results are robust to includ-
ing controls.
Table A2—Enforcement Effects.

The Impact of Ex Post Repayment Incentive within Ex Ante Treatment Group:
OLS with Controls for Referred Characteristics

Panel A. Ex ante approval incentive
Ex post repayment −0.100** −0.127** −0.120** −0.115**
(0.034) (0.045) (0.039) (0.046)
Mean in ex post 0.389 0.206 0.186 0.155
approval (0.064) (0.054) (0.054) (0.047)
Panel B. Ex ante repayment incentive

Ex post repayment −0.312** −0.072* −0.066 −0.098*
(0.095) (0.033) (0.040) (0.046)
Mean in ex post 0.519 0.226 0.256 0.189
approval (0.069) (0.058) (0.076) (0.054)
Application Score Post May 2009, ITC Score, Job Type, Requested Loan Amount, Requested
Term, Branch, Application Month, Application Year. All controls are for referrer characteris-
tics. Categorical variables are entered as fixed effects.
Finally, we again pool the data and assume that the enforcement and selection
effects are independent of each other. That is, we run the regression
yi = α + β 1 enforcei + β 2 selecti + β 3 Xi + ϵi ,
where Xi is a set of controls. In this case we can only control for referrer characteristics
as once again the referred characteristics are endogenous. Table A4 contains the results,
which do not differ significantly from those reported in Table 6 without controls.
Appendix B. Do Referrers Have Information?
The discussion in the text shows what is identified by the experiment. In this sec-
tion we argue that we can use the experiment to understand whether referrers have
information regarding repayment types. This is a more general question than asking
whether there is selection on repayment type. For example, if we see no selection
on repayment type it may be that referrers have information, but the experiment
Table A3—Selection Effects.

The Impact of Ex Ante Repayment Incentive within Ex Post Treatment Group:
OLS with Controls

Panel A. Ex post approval incentive
Ex ante approval Left out Left out Left out Left out
Ex ante repayment 0.046 0.046 0.035 0.027
(0.041) (0.035) (0.041) (0.031)
Mean in ex ante 0.389 0.206 0.186 0.155
approval (0.064) (0.054) (0.054) (0.047)
Panel B. Ex post repayment incentive

Ex ante approval Left out Left out Left out Left out
Ex ante repayment 0.007 0.009 0.018 0.028
(0.100) (0.080) (0.075) (0.063)
Mean in ex ante 0.258 0.095 0.076 0.047
approval (0.054) (0.037) (0.039) (0.027)
Notes: Penalty interest is charged by the lender if a borrower is late in making an expected pay-
ment. A loan is charged off if the lender deems that there is no probability that it will be repaid.
Standard errors are in parentheses. Ex ante incentive is the incentive that the referrer faced
when choosing a friend to refer. Ex post incentive is the incentive that the referrer faced after
the loan had been approved. Approval implies the loan had to be approved in order to earn the
bonus, and repayment implies the loan had to be repaid in order to earn the bonus. Controls:
Female, Age, Disposable Income, Salary Occurrence, Education, Application Score, Job Type,
Requested Loan Amount, Requested Term, Branch, Application Month, Application Year. All
controls are for referrer characteristics. Categorical variables are entered as fixed effects.
did not lead them to reveal that information. We will argue in this section and the
empirical work that follows that the experiment did lead referrers to attempt to select
high repayment types, but that this selection was (subject to power considerations)
unsuccessful. We interpret this as evidence that referrers in fact have no useful infor-
mation—at least that can be extracted with incentives of the size we use. From a
contract design perspective, this means that our experiment can be useful in ruling
out a whole class of mechanisms, rather than just implying that our particular exam-
ple is not effective in inducing selection.
DEFINITION 3 [Referrer Knowledge]: We say that referrers have information about

repayment type if
(B1) E(θ | θˆ = θˆ H & app) > E(θ | θˆ = θˆ L & app).
As with the definitions of selection in the main text, this statement is conditional
on loan approval: we are interested in whether referrers have information that pre-
dicts repayment above and beyond the information that the lender observes.
Table A4—Pooled Impact of Selection and Enforcement Treatments

on Key Outcome Variables: OLS with Controls (Same as Table 6 but with controls)
Outcome measure Penalty interest Not paid on time Portion owing Loan charged off
Enforcement −0.168*** −0.117** −0.130** −0.109**
(0.065) (0.055) (0.054) (0.047)
Selection −0.009 0.021 0.018 0.032
(0.074) (0.061) (0.062) (0.053)
Mean in 0.389 0.206 0.186 0.155
left out (0.064) (0.054) (0.054) (0.047)
ITC Score, Job Type, Requested Loan Amount, Requested Term, Branch, Application Month,
Application Year. All controls are for referrer characteristics. Categorical variables are entered
as fixed effects.
Inferring whether referrers have information can, but need not, be straightfor-
ward. If S(A) > 0 it must be that the referrers have information about repayment
type. But if S(A) ≤ 0 it is not immediately clear whether referrers have informa-
tion. It may be that referrers have information, but that they have not been induced
to reveal it. For example, our model allows for the possibility that referrers believe
approval types γˆ and repayment types θˆ to be highly correlated. In this case there
may be no differential incentive to refer high types in the ex ante repayment treat-
ment. Alternatively, because BA ≥ BR, referrers in the ex ante approval treatment
may be willing to pay a higher cost λ to make a referral. If λ is correlated with
repayment type we may even see S(A) < 0 , but this would not necessarily show
a lack of information. Finally, it could just be that referrers tend to only have one
potential referred.
We argue that if E(λ | r) > E(λ | a) , S(A) = 0 , S(R) = S(A), and σˆ is positively
correlated with σ , then we can infer that referrers do not have information. To see
this, note that we can write
S(A) ≡ Pr (θˆ = θˆ H | r)E(θ | θˆ H & app)
+ (1 − Pr (θˆ = θˆ H | r))E(θ | θˆ L & app)
− (Pr(θˆ = θˆ H | a)E(θ | θˆ H & app)
+ (1 − Pr (θˆ = θˆ H | a))E(θ | θˆ L & app)).

Consequently, if the experiment induces selection in the sense that
(SELECTION) Pr (θˆ = θˆ H | r) > Pr (θˆ = θˆ H | a),
then S(A) = 0 implies that referrers have no information. To complete the argu-
ment we need to show that E(λ | r) > E(λ | a) and S(R) − S(A) = 0 imply
SELECTION.
The condition E(λ | r) > E(λ | a), combined with (1) and (2), implies that
(B2) E(γˆ (BR(θˆ + σe ∗(σˆ )) − c(e ∗(σˆ ))) | r)
> E(γˆ (BR(θˆ + σe ∗(σˆ )) − c(e ∗(σˆ ))) | a) .
This follows because there must be some expected return to making a high cost
referral. Condition (B2), combined with the assumption that c is strictly con-
vex, then implies that E(θˆ + σe ∗(σ) | r) > E(θˆ + σe ∗(σ) | a): that the ex ante
incentive induced selection on the probability of repayment conditional on antic-
ipated effort. To complete the argument, observe that if S(R) = S(A), then
E(θˆ + σe ∗(σ) | r) > E(θˆ + σe ∗(σ) | a) implies E(θˆ | r) > E(θˆ | a) so long as we
are willing to make the weak assumption that σˆ is positively correlated with σ: that
those perceived to be more malleable are, in reality, more malleable.
Finally, it is worth noting that our restriction to two repayment types may be
restrictive in the case where S(A) = 0. Our experiment provides a limited incen-
tive (BR) to refer a good repayment type. It may be that this incentive induces
SELECTION within a set of low-cost referreds, and that a larger incentive would
induce selection across a larger group of potential referreds. For example, it may
be that the referrers know people who are high repayment types, but that it is very
costly to contact them and make the referral.
REFERENCES
Ahlin, Christian, and Robert M. Townsend. 2007. “Using Repayment Data to Test Across Models of
Joint Liability Lending.” Economic Journal 117 (517): F11–51.
Armendáriz, Beatriz, and Jonathan Morduch. 2010. The Economics of Microfinance. Cambridge:
MIT Press.
Banerjee, Abhijit Vinayak. 2013. “Microcredit Under the Microscope: What Have We Learned in the
Past Two Decades, and What Do We Need to Know?” Annual Review of Economics 5: 487–19.
Banerjee, Abhijit V., Timothy Besley, and Timothy W. Guinnane. 1994. “Thy Neighbor’s Keeper: The
Design of a Credit Cooperative with Theory and a Test.” Quarterly Journal of Economics 109 (2):
491–515.
Banerjee, Abhijit V., and Esther Duflo. 2014. “Do Firms Want to Borrow More? Testing Credit Con-
straints Using a Directed Lending Program.” Review of Economic Studies 81 (2): 572–607.
Banerjee, Abhijit V., and Andrew F. Newman. 1993. “Occupational Choice and the Process of Devel-
opment.” Journal of Political Economy 101 (2): 274–98.
Beaman, Lori, and Jeremy Magruder. 2012. “Who Gets the Job Referral? Evidence from a Social Net-
works Experiment.” American Economic Review 102 (7): 3574–93.
Benabou, Roland, and Jean Tirole. 2003. “Intrinsic and Extrinsic Motivation.” Review of Economic
Studies 70 (3): 489–520.
Besley, Timothy, and Stephen Coate. 1995. “Group lending, repayment incentives and social collat-
eral.” Journal of Development Economics 46 (1): 1–18.
Besley, Timothy, and Maitreesh Ghatak. 2005. “Competition and Incentives with Motivated Agents.”
American Economic Review 95 (3): 616–36.
Bond, Philip, and Ashok S. Rai. 2008. “Cosigned vs. group loans.” Journal of Development Econom-
ics 85 (1–2): 58–80.
Bryan, Gharad, Dean Karlan, and Jonathan Zinman. 2015. “Referrals: Peer Screening and Enforce-
ment in a Consumer Credit Field Experiment: Dataset.” American Economic Journal: Microeco-
nomics. http://dx.doi.org/10.1257/mic.20130234.
Carpena, Fenella, Shawn Cole, Jeremy Shapiro, and Bilal Zia. 2013. “Liability Structure in Small-Scale
Finance: Evidence from a Natural Experiment.” World Bank Economic Review 27 (3): 437–69.
de Janvry, Alain, Craig McIntosh, and Elisabeth Sadoulet. 2010. “The supply-and demand-side
impacts of credit market information.” Journal of Development Economics 93 (2): 173–88.
de Mel, Suresh, David McKenzie, and Christopher Woodruff. 2008. “Returns to Capital in Microen-
terprises: Evidence from a Field Experiment.” Quarterly Journal of Economics 123 (4): 1329–72.
Einav, Liran, Amy Finkelstein, Stephen P. Ryan, Paul Schrimpf, and Mark R. Cullen. 2013.
“Selection on Moral Hazard in Health Insurance.” American Economic Review 103 (1): 178–219.
Fischer, Greg. 2013. “Contract Structure, Risk-Sharing, and Investment Choice.” Econometrica 81 (3):
883–939.
Galor, Oded, and Joseph Zeira. 1993. “Income Distribution and Macroeconomics.” Review of Eco-
nomic Studies 60 (1): 35–52.
Ghatak, Maitreesh. 1999. “Group lending, local information and peer selection.” Journal of Develop-
ment Economics 60 (1): 27–50.
Ghatak, Maitreesh, and Timothy W. Guinnane. 1999. “The economics of lending with joint liability:
Theory and practice.” Journal of Development Economics 60 (1): 195–228.
Giné, Xavier, Jessica Goldberg, and Dean Yang. 2012. “Credit Market Consequences of Improved
Personal Identification: Field Experimental Evidence from Malawi.” American Economic
Review 102 (6): 2923–54.
Giné, Xavier, and Dean S. Karlan. 2014. “Group versus individual liability: Short and long term evidence
from Philippine microcredit lending groups.” Journal of Development Economics 107: 65–83.
Gneezy, Uri, and Aldo Rustichini. 2000. “A Fine Is a Price.” Journal of Legal Studies 29 (1): 1–18.
Gunnsteinsson, Snaebjorn. 2014. “Information Asymmetries in Crop Insurance.” http://www.
snaebjorn.com/uploads/1/2/2/7/12277061/gunnsteinsson__information_asymmetries_in_crop_
insurance.pdf.
Karlan, Dean, Melanie Morten, and Jonathan Zinman. 2012. “A Personal Touch: Text Messaging for
Loan Repayment.” National Bureau of Economic Research (NBER) Working Paper 17952.
Karlan, Dean, and Jonathan Zinman. 2009. “Observing Unobservables: Identifying Information
Asymmetries With a Consumer Credit Field Experiment.” Econometrica 77 (6): 1993–2008.
Karlan, Dean, and Jonathan Zinman. 2010. “Expanding Credit Access: Using Randomized Supply
Decisions to Estimate the Impacts.” Review of Financial Studies 23 (1): 433–64.
Klonner, Stefan, and Ashok S. Rai. 2010. “Cosigners as Collateral.” Journal of Development Econom-
ics 60 (1): 27–50.
Rai, Ashok S., and Tomas Sjöström. 2004. “Is Grameen Lending Efficient? Repayment Incentives and
Insurance in Village Economies.” Review of Economic Studies 71 (1): 217–34.
Stiglitz, Joseph E. 1990. “Peer Monitoring and Credit Markets.” World Bank Economic Review
4 (3): 351–66.
Stiglitz, Joseph E., and Andrew Weiss. 1981. “Credit Rationing in Markets with Imperfect Informa-
tion.” American Economic Review 71 (3): 393–410.
Varian, Hal R. 1990. “Monitoring Agents With Other Agents.” Journal of Institutional and Theoretical
Economics 146 (1): 153–74.

2015 - Referrals Peer Screening and Enforcement in Consumer Credit

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2015 - Referrals Peer Screening and Enforcement in Consumer Credit

Uploaded by

Copyright:

Available Formats

American Economic Journal: Microeconomics 2015, 7(3): 174–204

Referrals: Peer Screening and Enforcement

Empirical evidence on peer intermediation lags behind both theory

Ex ante incentive: Bonus promised

Figure 1. 2 × 2 Experimental Design—Repayment Incentives

Results from our demonstration show strong enforcement effects. By comparing

I. Market and Lender Overview

II. The Experiment

If the friend is approved for a loan

If the friend successfully repays a loan

Figure 2. Referral Cards

Appendix B analyzes the extent to which the experiment allows us to determine

A. The Referral Decision

A market consists of N potential borrowers (referreds), each of whom is charac-

(i) θ ∈ {θ L, θ H} : a repayment type;

(ii) σ ∈ {σ L, σ H} : a malleability type;

Figure 3. Venn Diagram Showing the Distribution of Potential Referreds

We assume that {θ H, θ L} and {σ H, σ L} are such that the probability of repay-

max {BR(θˆ + σˆ e) − c(e)},

(1) Uj(A) ≡ max {γˆi BA − λij},

Uj(R) ≡ max {γˆi (BR(θˆ + σˆ e ∗(σˆ )) − c(e ∗(σˆ ))) − λij} .

B. Map from Model to Data

S(A) = E(θ | rA & app) − E(θ | aA & app),

We also have two potential measures of enforcement. We define

N(r) = E(θ + σe ∗(σˆ ) | rR & app) − E(θ | rA & app)

as the difference in repayment rates across enforcement treatments conditional on

N(a) = E(θ + σe ∗(σˆ ) | aR & app) − E(θ | aA & app).

C. Identifying Selection and Enforcement

The identification of enforcement effects is straightforward if the ex post treatment

DEFINITION 1 [Selection on Repayment Type]: We say that our repayment incen-

E(θ | r & app) > E(θ | a & app).

In terms of Figure 3, the probability of repayment conditional on approval is pro-

DEFINITION 2 [Selection on Malleability]: We say that our incentive induced pos-

In terms of Figure 3 we say that there is selection on malleability if YZ/WX

E(σ | r & app) > (<)E(σ | a & app).

Therefore, S(R) − S(A) > (<)0 implies positive (negative) selection on

Table 1 provides a summary of the characteristics of potential referrers over the

Table 1—Demographic Variables of All Borrowers during Experiment

B. Integrity of the Randomization

Opportunity handed out 4,408 referral cards to potential referrers—borrowers

Table 2—Testing the Balance of Referrer Characteristics

Ex ante incentive Approval Repayment

F-test of joint significance 0.560 0.930 0.810 0.440

Observations 4,408 4,408 4,408 4,408

A. Who Refers and Who Gets Referred?

Among eligible clients, 430 of 4,408, or 10 percent, made a referral. Making a

Table 3—Testing the Balance of Referred Characteristics

Whole App. score Before May

F-test of joint significance 0.810 1.640 1.150 0.990

Observations 123 120 120 108

Table 4—The Impact of Ex Ante Treatment Group on Referred Characteristics:

Same church Coworker

Constant −1.932*** 0.686***

B. Measuring Repayment Performance

We identify screening and enforcement effects by comparing the repayment

Table 5—Key Outcome Variables: Mean Differences across Treatment Groups

Ex ante incentive Ex ante incentive

E. Improving Precision and Checking Robustness

Under the assumption that malleability is uncorrelated with repayment type,

VI. Alternative Explanations and Other Considerations

In this section we discuss alternative interpretations of the results.

In theory, the enforcement effect could be driven by side-payments from the

Observations 245 240 240 239

D. Reputation and Other Considerations

Constant −1.932* 0.686*